Dynamic Gesture Recognition Model Based on Millimeter-Wave Radar With ResNet-18 and LSTM

Area of Science:

• Computer Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Millimeter-wave radar, like Soli radar, offers high positional accuracy for recognizing subtle movements, crucial for advancing Human-Computer Interaction (HCI).
• Dynamic gesture recognition remains a challenging area, requiring models capable of processing complex spatio-temporal data effectively.

Purpose of the Study:

• To propose and validate a novel deep learning model for dynamic gesture recognition using millimeter-wave radar signals.
• To leverage the strengths of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) for feature extraction and sequence modeling.

Main Methods:

• A multi-layer model combining ResNet-18 for spatial feature extraction and Long Short-Term Memory Networks (LSTM) for temporal feature extraction was developed.
• Input data consisted of velocity-range Doppler images derived from raw millimeter-wave radar signals.
• The model was trained and evaluated on the Soli dataset, specifically designed for dynamic gesture recognition.

Main Results:

• The proposed ResNet-18 and LSTM model achieved a high accuracy of 92.55% in dynamic gesture recognition.
• The model effectively addressed gradient issues in deep networks using ResNet-18 and captured long-range temporal dependencies with LSTM.
• Experimental comparisons demonstrated superior performance compared to traditional dynamic gesture recognition methods.

Conclusions:

• The developed deep learning model demonstrates significant effectiveness and high accuracy for dynamic gesture recognition using millimeter-wave radar.
• The integration of ResNet-18 and LSTM provides a robust approach for handling the spatio-temporal complexities inherent in radar-based gesture data.
• This research validates the potential of advanced deep learning architectures for enhancing HCI through precise gesture interpretation.